alphacep / vosk-api

Offline speech recognition API for Android, iOS, Raspberry Pi and servers with Python, Java, C# and Node
Apache License 2.0
7.35k stars 1.04k forks source link

Feature Request: Number only model #1598

Open nduc opened 1 week ago

nduc commented 1 week ago

Hi, Is it possible to get a number only model? I'm working on a voice calculator and converter for people working in environment where their hands are either busy or dirty or unable to use keypad or touch screen. I'm making an app where they can speak the numbers and operations. Including fractions and decimals. The current model is a bit poor on number recognition. For example, we can say 2023 as "two zero two three" or "two oh two three" or "twenty twenty three" or "two thousand twenty three". I have code to convert the words to number. The model struggles with some numbers so I'm getting things like nineteen as "night team". I figure a number focused model will help the recognition and eliminate many false recognitions. Thank you!

nshmyrev commented 1 week ago

You can adapt it yourself. See

https://github.com/alphacep/vosk-api/blob/master/python/example/colab/vosk-adaptation.ipynb

https://alphacephei.com/vosk/lm

nduc commented 1 week ago

Thank you!